Jina AI
Jina AI develops models that find relevant content across text, images and other media. Han Xiao founded the company in 2020. Elastic acquired it in October 2025, adding Jina's embeddings, rerankers and reading models to its search business.
Company & leadership
- Origin
- Berlin, Germany
- Owner
- Elastic acquired 100% of Conic AI Technology Limited and its subsidiaries on 7 October 2025
- Founder's role at acquisition
- Han Xiao became Vice President of AI at Elastic
- Acquisition consideration
- US$43.4 million, including a US$6.9 million indemnity holdback; reported in the FY2026 annual filing
Founder
- Han Xiao
Company and ownership
Jina AI began in Berlin in 2020 and developed a specialist search-model business. The company's own history identifies Han Xiao as its founder. Its models focus on retrieving, ordering and extracting information, with generation used for selected reading and vision tasks.
Elastic's FY2026 filing dates the purchase of all shares in Conic AI Technology Limited and its subsidiaries, collectively Jina AI, to 7 October 2025. The public announcement followed on 9 October. The filing reports US$43.4 million in consideration. This is an acquisition price, not a current standalone market value.
At the acquisition announcement, former Jina CEO Han Xiao became Vice President of AI at Elastic. Jina continues as a model-development brand within Elastic. The announced role change is documented here without assigning an unverified current standalone Jina CEO title.
Search models and trained derivatives
The model families solve different parts of search. Embeddings encode content into vectors. ColBERT retains token-level vectors for later matching. Rerankers examine a query and candidate results more closely. ReaderLM transforms HTML, while Jina OCR reads rendered pages. The catalog keeps those trained models separate from the Reader, Embeddings and Reranker services that deliver them.
Jina uses both its own encoder designs and further training of external backbones. Examples include Qwen-based code embeddings and rankers, a EuroBERT-based v5 Nano encoder, and the DeepSeek-OCR-based document parser. Their model entries identify the upstream model; trained adaptations are not presented as original backbone inventions.
Releases and access
The listed releases include older downloadable models and documented code fine-tunes. Task adapters, smaller vector widths, quantized files and runtime conversions are described within their parent model entries. They do not each create another independently trained model. A superseded release can still have accessible weights.
Jina's public repository list also contains implementation packages, debugging copies, an untrained model and segmentation adapters with incomplete template cards. Those items are not promoted here as documented customer models. Planned v1 embedding sizes without a verified release are likewise excluded. This is a source-backed inventory, not a claim to list unpublished internal experiments.
Access terms differ by generation. Several early encoders and rankers use Apache 2.0. Many newer downloads use CC BY-NC 4.0, and Embeddings v4 explicitly uses the Qwen Research License. A managed API offer does not change the license of downloaded weights.